Natera's voice agent on Bedrock AgentCore books phlebotomy appointments at 6.8s median latency and under a cent a call

AWS Machine Learning Blog·medium signalXBlueskyLinkedInCopy link

The 2026-08-26 write-up details a dual-WebSocket bridge that separates telephony streaming from model inference, with one connection to Twilio for the patient call and a second to voice processing, and the AgentCore-hosted agent sitting between them intercepting tool requests and injecting results back into the conversation. Claude Haiku generates filler acknowledgments to cover tool-execution gaps. Four weeks of production data covered 4,744 calls, up 5.5% versus the prior system, with 100% tool-calling accuracy over 500 end-to-end simulations, over 90% response accuracy on scheduling outcomes, and short-call abandonment down from 22% to 12%.